A novel deep learning-based grading system for assessing breast arterial calcification on mammograms, as an independent risk factor for predicting adverse cardiovascular events

PURPOSE: To develop a deep learning-based framework for automated detection and grading of breast arterial calcification (BAC) on mammograms, and to evaluate its association with major adverse cardiovascular events (MACE). MATERIAL AND METHODS: This retrospective case-control study used two datasets: SegModel (1270 mammograms with BAC annotations) and MACEPred (3190 mammographic cases of women with MACE, plus 6458 controls). A U-Net segmentation model and Cox proportional hazards models were used to assess adjusted hazard ratios (HRs) for four BAC grading strategies: binary categorization, area-based grading (none, mild, moderate, severe), intensity-based grading, and a combined approach. FINDINGS: The U-Net attained a Jaccard similarity coefficient of 0.582, accuracy of 0.996, precision of 0.801, F1 score of 0.756, and recall of 0.716 in segmenting BAC. The presence of BAC was associated with an adjusted HR of 1.16 (95% CI 1.10-1.23). For area-based grading, the HRs for mild, moderate, and severe grades were 1.13 (95% CI 1.06-1.20), 1.30 (95% CI 1.18-1.45), and 1.58 (95% CI 1.15-2.18), respectively. Intensity-based grading showed respective HRs of 1.08 (95% CI 1.00-1.17), 1.10 (95% CI 1.01-1.20), and 1.18 (95% CI 1.09-1.28). The combined approach demonstrated respective HRs of 1.102 (95% CI 1.04-1.17), 1.249 (95% CI 1.14-1.37), and 1.654 (95% CI 1.30-2.10) for mild, moderate, and severe grades. CONCLUSION: Our study presents a novel automated framework for BAC assessment that provides independent insights into cardiovascular risk in women. Combined area and intensity grading reflected increasing MACE risk across BAC severity levels, although its incremental predictive improvement was limited.